Optimal Selection of Airport Runway Conﬁgurationsdbertsim/papers...Bertsimas, Frankovich, and Odoni: Optimal Selection of Airport Runway Conﬁgurations 1408 Operations Research

OPERATIONS RESEARCHVol. 59, No. 6, November–December 2011, pp. 1407–1419issn 0030-364X �eissn 1526-5463 �11 �5906 �1407 http://dx.doi.org/10.1287/opre.1110.0956

© 2011 INFORMS

Optimal Selection of Airport Runway Configurations

Dimitris Bertsimas, Michael Frankovich, Amedeo OdoniOperations Research Center, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139

{[email protected], [email protected], [email protected]}

We present a mixed integer programming (MIP) model to solve the problems of (i) selecting an airport’s optimal sequenceof runway configurations and (ii) determining the optimal balance of arrivals and departures to be served at any moment.These problems, the runway configuration management (RCM) problem and the arrival/departure runway balancing (ADRB)problem, respectively, are of critical importance in minimizing the delay of both in-flight and on-the-ground aircraft alongwith their associated costs. We show that under mild assumptions on the time required to change between configurations,large realistic problem instances can be solved within several seconds. Furthermore, as assumptions are relaxed, optimalsolutions are still found within several minutes. Comparison with a sophisticated baseline heuristic reveals that in manycases the potential reduction in cost from using the method is significant and could be expected to be of the order of atleast 10%. Finally, we present an extension of the MIP model to solve these two problems for a group of airports in ametropolitan area such as New York (metroplex), where operations at each airport within the metroplex might have animpact on operations at some of the other airports due to limitations in shared airspace.

Subject classifications : transportation: air traffic; programming: integer; applications.Area of review : Transportation.History : Received November 2009; revisions received August 2010, January 2011; accepted April 2011. Published online

in Articles in Advance November 18, 2011.

IntroductionThe combination of runways that are active at any particulartime at an airport is known as the “runway configuration”in which the airport operates at that time. The sequence ofconfigurations selected by controllers greatly influences anairport’s capacity to serve demand for arrivals and depar-tures. This paper is concerned with determining a sequenceof runway configurations and the assignment of arrivals anddepartures to the active runways that together minimize thecost incurred at the airport due to delayed aircraft.

Figure 1 shows a schematic representation of Boston’sLogan International Airport (BOS) with its six runways.The runway configuration shown in use in the figure con-sists of three active runways, shown in bold, with the otherthree being idle. The direction of operations on each isindicated. Two of the runways are used for both arrivalsand departures, while the third is used only for departures.For multi-runway airports, the number of possible configu-rations can be large. For instance, BOS typically employsmore than 20 different configurations during a year by mak-ing use of its six runways in different ways.

The availability of certain configurations at most majorairports might be restricted by the weather conditions pre-vailing at any particular time. For example, a runwaycannot be operated in the presence of strong crosswinds(relative to the orientation of that runway); or a runwaymight not be adequately instrumented for operations underpoor visibility conditions. There are also physical limitsto the capacity of each configuration, i.e., to the number

of arrivals and departures that can be accommodated in agiven length of time while operating in a given configura-tion. In an operational context, capacity typically is mea-sured as the expected number of movements that can takeplace in the presence of continuous demand. This is alsoknown as the maximum throughput capacity and is mea-sured as the number of arrivals and departures per unit oftime, typically 10, 15, or 60 minutes. In making their deci-sions concerning the best runway configuration to use atany given time, controllers take into account the capacityof each available configuration, as well as the scheduleddemand for arrivals and departures, and the weather fore-cast, which influences the future availability of the differentrunway configurations.

In our model, we represent the capacity of different con-figurations through the runway configuration capacity enve-lope (RCCE), studied and used in Gilbo (1993, 1997), Hall(1999), and Gilbo and Howard (2000) and discussed indepth in de Neufville and Odoni (2003). An RCCE is aconcave piecewise linear function that defines the set of fea-sible operating points that can be achieved under a givenconfiguration and under given weather conditions. A run-way configuration typically will have more than one RCCE,each corresponding to different weather conditions. Figure 2shows two typical RCCE that could describe the capac-ity for arrivals and departures of a runway configurationlike that shown in Figure 1. Note that the horizontal andvertical axes show, respectively, the number of departuresand arrivals that are demanded or performed during theselected unit of time, be that 10, 15, or 60 minutes. The

1407

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

Bertsimas, Frankovich, and Odoni: Optimal Selection of Airport Runway Configurations1408 Operations Research 59(6), pp. 1407–1419, © 2011 INFORMS

Figure 1. A runway configuration used at BOS, withthe corresponding active runways shown inbold.

outermost RCCE shows the capacity available under visualmeteorological conditions (VMC), while the inner one cor-responds to the same configuration under instrument condi-tions (IMC), when visibility is limited. Any integral point onor within the available RCCE corresponds to a feasible oper-ating point, and any point outside is infeasible. For example,Point 4 is feasible in VMC but infeasible in IMC. In the lat-ter case, the configuration does not have sufficient capacityto accommodate simultaneously the number of arrivals anddepartures associated with Point 4 during a single unit oftime, resulting in the queueing of aircraft. The total numberof RCCE that need to be considered, assuming the weatherconditions to be known, is equal to the number of configu-rations, e.g., about 20 in the case of BOS.

In this paper, we present mixed integer programmingmodels (MIP) that solve two problems simultaneously.The first, the runway configuration management (RCM)

Figure 2. Runway configuration capacity envelopes(RCCE) for a single configuration underVMC and IMC.

2 1

3

00

2

4

6

8

10

12

14

16

Arr

ival

s (p

er ti

me

inte

rval

)

18VMC

IMC

2 4 6

Departures (per time interval)8 10 12 14

4

problem, is to find a schedule of runway configurations thatwill best serve the demand over a specified time horizon.This is the problem we have outlined thus far. Then, giventhis schedule of configurations, there remains a degree offreedom as to the exact balance of arrivals versus departuresto serve during any time interval. Determining the optimalarrival/departure runway balancing (ADRB) is the secondproblem controllers face. Gilbo (1997) proposed an integerprogram for the ADRB to minimize queueing by control-ling the utilization of near-runway airspace and coordinat-ing operations at arrival and departure fixes subject to agiven schedule of configuration use. Furthermore, Gilbo andHoward (2000) proposed an integer program to solve theADRB problem assuming a constant overall airport RCCE.As far as we are aware, we present here the first optimiza-tion approach to the combined RCM and ADRB problems.

In practice, these problems are solved today largely onthe basis of experience. Under this essentially “manual”approach, the FAA examines the weather forecast and theexpected demand for the next several hours at each majorairport and determines the sequence of runway configu-rations to be used, including the assignment of arrivalsand departures to runways for each selected configuration.In doing so, location-specific issues, such as constraintson runway use imposed by noise or other considerations(see §4) are also taken into account. The difficult and mostimportant instances occur when the weather is highly vari-able or poor (e.g., changing wind directions, occurrenceof IMC). In such cases, the combination of changes inrunway configurations—with attendant partial or full lossof capacity during changeover times—and of the need toselect among “low” RCCE is likely to result in a significantbacklog of arrivals and/or departures from one time periodto the next. Under such conditions, the situation becomeshighly dynamic, and sequences and assignments might bemodified frequently in response. The models described herecan be most helpful in precisely these cases.

The solution to the RCM and ADRB problems might alsoimpact air traffic flow management (ATFM) actions that theFAA and airlines take in the context of the collaborativedecision making (CDM) program. The solution essentiallydetermines the “airport acceptance rate” (AAR) that ATFMuses to specify the number of arrivals that can be sched-uled at each airport in each time period, i.e., the numberof arrival “slots” that will be allocated under CDM amongthe airlines using the airport. If the AAR at an airport islow compared to the forecast demand for arrivals there, aground delay program (GDP) can be initiated by ATFM forthat airport, thus delaying before take-off (at their originat-ing airport) flights headed to that airport. In this sense, theRCM and ADRB problems have implications not only atthe local level but also at a national one.

In solving these problems, our objective is to mini-mize the total weighted cost of delays to queued aircraft,where the weights might take into account the cost of fuelburn on the airport’s surface (for departures) and in the

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

Bertsimas, Frankovich, and Odoni: Optimal Selection of Airport Runway ConfigurationsOperations Research 59(6), pp. 1407–1419, © 2011 INFORMS 1409

near-airport airspace (for arrivals), as well as other con-siderations (e.g., controller workload) that might result inarrival delays being assigned a higher weight than delaysto departures. Note that (i) the changeover times neededto transition between successive runway configurations and(ii) the restricted availability of certain configurations atcertain times due to weather conditions are both criticalaspects of the optimization to be performed. The formerprecludes the optimality of a greedy policy that simplyoperates in the best configuration for the demand in thecurrent time interval (without taking into account the futuresequence of configurations), while the latter requires thatwe consider the “lifespan” of each candidate configuration.

In our models, selecting a configuration is equivalent toselecting the outermost available RCCE corresponding tothat configuration, this RCCE being determined by currentweather conditions. (We assume that, for a given config-uration, the IMC RCCE is always fully contained withinthe one for VMC.) Selecting the arrival/departure balancelevel is then equivalent to selecting an integral point on orwithin this RCCE. This selection also affects how delay,if any, is allocated between arrivals and departures. Forinstance, in Figure 2, let Point 1 correspond to the demand4d11 a15 during hour 1 and assume that the airport oper-ates in VMC. Because Point 1 lies outside the availableRCCE, some of the demand for the hour in question can-not be accommodated. In this particular instance, an optimalchoice of operating point, considering only this hour, clearlylies somewhere on the line connecting Points 2 and 3.

We assume that the break points of each RCCE areintegral so that, along with nonnegativity constraints, theRCCE defines the convex hull of its feasible operatingpoints in the two-dimensional departures-arrivals operat-ing space. These RCCE can be obtained by observing pastoperating points and fitting minimal concave piecewise lin-ear functions which bound these points (see Gilbo 1993and Kellner and Kösters 2008). Alternatively, computa-tional models of airport operations can be used to generatethem, as discussed in Swedish (1981), de Neufville andOdoni (2003), and Stamatopoulos et al. (2004).

Contributions and Outline of Paper

The most important contribution of this paper is the devel-opment of mixed integer programming models to solve theRCM and ADRB problems efficiently for a single airport,as well as the proposed extension of the models to themetroplex case. In particular:

• In §1, we present a MIP for the single airport problem,assuming that the changeover time between configurationsis a constant, C, for all pairs of configurations. While thisMIP does not define the convex hull of feasible solutions,several computational experiments on large realistic datasets in §1.3 indicate that the formulation is a strong one.Furthermore, the results of these experiments indicate thatthe potential benefit of using this approach is significantand that improvements of the order of at least 10% could

be expected relative to a sophisticated optimization-basedheuristic.

• In §2, we relax the above assumption to allow thechangeover time between any two configurations to belongto a set of the form 8iC2 i = 11 0 0 0 1 �9, where C is a con-stant. Computational experiments show that for � ∈ 82139,optimal solutions are still obtained within several minutes.

• In §3, we present an extension of the model to opti-mize over a metroplex of airports.

• In §4, we show that important environmental con-straints can be easily incorporated into the model.

1. A Mixed Integer Programming Modelfor a Single Airport with ConstantChangeover Times

Here we seek to solve the RCM and ADRB problems, out-lined in the introduction, for a single airport. We assumethat the time required to change between configurations isa constant, C, and that the relevant data, as listed in thedefinitions below, are available and known with certainty.The latter assumption is reasonable given that we considerin our model a short time horizon of a few hours. Further-more, we envisage the method being implemented using arolling horizon approach, re-solving every 30 minutes or soto account for any divergence of observed from expecteddata.

T= the set of time intervals, 81121 0 0 0 1 T 9;Kt = the set of configurations available during time inter-

val t;Jt

k = the set of linear pieces of the outermost RCCE avail-able for configuration k at time t;

at = the number of arrivals scheduled for time interval t;dt = the number of departures scheduled for time

interval t;ct = the cost of delaying a single arrival for one time inter-

val at time interval t;qt = the cost of delaying a single departure for one time

interval at time interval t.

In addition, we let the jth piece of the outermost RCCEavailable for configuration k at time t, an affine function,be defined by the parameters �jk > 0, �jk ¶ 0, and �jk ¾ 0.While these parameters depend on t, because j belongs toa set that depends on t, we omit this from our notation.

Note that the time horizon is discretized such that thelength of one time interval is equal to the changeovertime C, which is typically of the order of 10 minutes.In this way, we can model a changeover by allowing noarrivals or departures to be served during the time intervalat which the changeover occurs. Moreover, the set of timeintervals, T, includes a number of time intervals at the endof the original time horizon, with no scheduled arrivals ordepartures. During these added time intervals, any queuesthat have built up can be cleared. We assume that we are

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


able to clear all queues without having to cancel or re-routeany flights.

Our decision variables are the following and are definedfor every t only for those configurations k that areavailable:

zkt =

11 if we operate in configuration k at time t1

01 otherwise3

ykt =

the number of arrivals served at time t1if we operate in configuration k at time t1

01 otherwise3

xkt =

the number of departures served at time t1if we operate in configuration k at time t1

01 otherwise3ut = the number of arrivals which go unserved at time t;vt = the number of departures which go unserved at

time t.

The mixed integer optimization problem is as follows:

P12 min∑

t∈T

4ctut + qtvt5

s/t0 ut − ut−1 +∑

k∈Kt

ykt = at1 ∀ t ∈T1 (1)

vt − vt−1 +∑

k∈Kt

xkt = dt1 ∀ t ∈T1 (2)

�jkykt −�jkxkt −�jkzkt ¶ 01

∀ j ∈Jtk1 ∀k ∈Kt1 ∀ t ∈T1 (3)

∑

k∈Kt

zkt ¶ 11 ∀ t ∈T1 (4)

∑

k′∈Kt−1\8k9

zk′1 t−1 + zkt ¶ 11

∀k ∈Kt1 ∀ t ∈T\8191 (5)

zkt ∈ 801191 ∀k ∈Kt1 ∀ t ∈T1 (6)

ut1 vt ∈�+1 ∀ t ∈T1 (7)

ykt ¾ 01 ∀k ∈Kt1 ∀ t ∈T1 (8)

xkt ¾ 01 ∀k ∈Kt1 ∀ t ∈T1 (9)

u01 v0 = 00 (10)

Our objective is to minimize the total cost of delaysincurred. Constraints (1), (2), (7), and (10) define the vari-ables ut and vt , the number of arrivals and departures,respectively, that go unserved at time t. Constraints (3)force the operating point, given that we operate in the kthconfiguration at time t, to lie within its RCCE, while forc-ing it to zero if zkt = 0. Constraints (4) state that at anytime t, we may operate in at most one configuration.

Constraints (5) invoke our fundamental assumption thatthe changeover time is equal to the length of one time inter-val. In this way, we model the cost of changing from one

configuration to another by enforcing a delay of one timeinterval. In other words, it is not possible to operate in con-figuration k at time t and also in configuration k′ 6= k attime t − 1.

One could also modify Constraints (5) to consider onlya subset of all pairs of configurations 8k1 k′9, thus model-ing the changeovers between the excluded pairs as beinginstantaneous. This would be suitable if the correspond-ing changeover times were very small compared to otherchangeovers.

Note that we have defined the variables ut and vt to beintegral in Constraints (7). Given our assumptions on thenature of the RCCE and the integrality of the arrival anddemand data, as well as Constraints (6) on zkt , it followsthat in an optimal solution to the resulting MIP, ykt and xktare integral. We therefore relax the integrality constraintsfor these two sets of variables and greatly reduce the num-ber of integer variables in the model.

We can then add two classes of valid inequalities tostrengthen the formulation. First, we add Inequalities (11)below, which are closely related to Constraints (5). Observethat in Inequalities (11), we require for all k that we cannotboth operate in configuration k at time t − 1 and also insome configuration k′ 6= k at time t, while in Constraints(5), we require for all k that we cannot both operate someconfiguration k′ 6= k at time t− 1 and in configuration k attime t.

zk1 t−1 +∑

k′∈Kt\8k9

zk′t ¶ 11 ∀k ∈Kt−11 ∀ t ∈T\8190 (11)

It is clear that while the set of feasible integer solutionsremains unchanged under this addition, we remove somenonintegral solutions from the solution set of the LP relax-ation of P1.

Now we generate for each time interval t a singleRCCE that defines the convex hull of the set 84x1 y5 ∈�2

+2

∃k ∈ Kt s.t. �jky − �jkx ¶ �jk1 ∀ j ∈ Jtk9, i.e., the mini-

mal piecewise linear concave envelope that majorizes allRCCE at time t. We let this RCCE be defined by theparameters �′

jt , �′jt , and � ′

jt , ∀ j ∈J′t . Then, observing that

yt4

=∑

k∈Ktykt and xt

4

=∑

k∈Ktxkt represent the number of

arrivals and departures served at time t, respectively, it isclear that 4xt1 yt5 must lie within this RCCE, and henceInequalities (12) below are valid.

� ′

jt

∑

k∈Kt

ykt −�′

jt

∑

k′∈Kt

xk′t ¶ �′

jt1 ∀ j ∈J′

t1 ∀ t ∈T0 (12)

Furthermore, these inequalities give the tightest possiblebound on the relation between feasible yt and xt , becausethe convex hull of this (in general, nonconvex) set has beendefined.

1.1. Example Problem

We now present a simple example to test the model andgain insight into its solution. We consider a time horizon

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Figure 3. Test Problem 1 data, with three configura-tions and 15 time intervals.

00

5

10

Sch

edul

ed a

rriv

als

in e

ach

time

perio

d

15

20

1

8

10

29

7 34

5

6

25

5 10

Scheduled departures in each time period15 20 25

Notes. RCCE 1 has a solid line, RCCE 2 a dotted line, and RCCE 3 adashed line. The scheduled demand is indicated for intervals 1 to 10 andis zero for intervals 11 to 15.

of 10 periods of demand, set T = 15 to allow time toserve all demand, and ct = 12, qt = 10, ∀ t ∈ T to capturethe typically greater cost of delaying arrivals compared todepartures. In all examples that follow, we shall use thesecost coefficients. The scheduled demand for arrivals anddepartures is displayed in Figure 3 along with the RCCEcorresponding to three configurations, which are availablethroughout the entire time horizon.

This problem, along with all others that follow in thispaper, was solved using AMPL CPLEX 11.2.10, using asingle thread, on a computer with an Intel(R) Core(TM) 2Duo E7400 Processor (2.80 GHz, 3 MB Cache, 1,066 MHzFSB) and 2 GB of RAM, running Ubuntu Linux. The opti-mal solution, which is displayed in more detail in Fig-ures 4, 5, and 6, was found in 0.04 seconds and consistsof operating in configuration 2 for intervals 1 and 2, con-figuration 3 for intervals 4 to 6, and then in configuration2 again from interval 8 onward.

Note that the “actual demand” in the system at time t,given the operating policy in the preceding time intervals,consists of ut−1 + at arrivals and vt−1 + dt departures, i.e.,any arrivals (departures) left in queue at the end of timeinterval t − 1 plus the scheduled arrivals (departures) intime interval t. We can then observe the following:

1. In Figure 4, the actual demand and scheduled demandare identical because the first demand point lies withinRCCE 2, with which we operate, and hence all demand isserved and there is no backlog added to the second timeinterval.

2. In Figure 5, because we are changing from configu-ration 2 to configuration 3, we have x3 = y3 = 0. In thisway, the third actual demand point is added to the fourthscheduled demand point to create the fourth actual demandpoint. Similar behavior is also seen with time intervals 7and 8 in Figure 6.

Figure 4. Test Problem 1 solution, intervals 1 and 2.

25201510

10

15

1

2

20

25

Demand for departures in each time period

Dem

and

for

arriv

als

in e

ach

time

perio

d

5

5

00

Notes. RCCE 1 has a solid line, RCCE 2 a dotted line, and RCCE 3 adashed line. Scheduled demand is indicated by a diamond, actual demandby a cross, although these are identical here.

3. When a change is made, it is made during an inter-val of relatively low actual demand and before a sequencethat is favored by the new configuration.

4. As time passes, the scheduled and actual demandcould diverge significantly.

In summary, we have learned that even for a very smallproblem, it is essential to account for the cumulative natureof demand and its dependence on our decisions; our earlydecisions might have a long-lasting impact on the overallproblem. These interactions might be very complex, andhence good decision-making will in general be difficultwithout the use of sophisticated tools.

1.2. A Baseline Policy

Before performing computational experiments on problemsof realistic size, we now devote our attention to presenting

Figure 5. Test Problem 1 solution, intervals 3 to 6.

25201510


5

34

6

6

5

4

0

10

15

20

25

Dem

and

for

arriv

als

in e

ach

time

perio

d

5

0

5

Notes. RCCE 1 has a solid line, RCCE 2 a dotted line, and RCCE 3 adashed line. Scheduled demand is indicated by a diamond, actual demandby a cross.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Figure 6. Test Problem 1 solution, intervals 7 to 12.

25201510

10

15

8

8

7

7

111211,12

10 10

99

20

25


Dem

and

for

arriv

als

in e

ach

time

perio

d

5

5

00

Notes. RCCE 1 has a solid line, RCCE 2 a dotted line, and RCCE 3 adashed line. Scheduled demand is indicated by a diamond, actual demandby a cross.

a baseline policy in order to obtain a reasonable indica-tion of the relative quality of the solutions obtained fromP1 compared with, say, current practice. The baseline pol-icy was designed (i) to obtain an estimate of a lowerbound on the improvement that could be observed in prac-tice by implementing the policy obtained from solving P1and (ii) to mimic the approach that might be taken by ahighly skilled controller. The method is developed througha “smart” heuristic, summarized below first at a high leveland then in detail.

1. Among all configurations that are available for a sig-nificant period of time starting now, choose the one that isbest (per the criteria described below) to operate in.

2. Operate in this configuration until the first time periodat which it is no longer available for use.

3. When that happens, observe the new state of the sys-tem: if there is no demand left, then stop; otherwise, returnto Step 1.

The algorithm is essentially a greedy algorithm, modifiedto avoid a large number of costly changeovers. We believethis to be a good approximation of controllers’ actions:given the difficulty of foreseeing the effect of one’s deci-sions on future demand, one cannot plan “manually” a goodconfiguration sequence too far into the future.

To choose which configuration is selected as “best,” wefirst restrict our list of eligible configurations to those thatare available for a reasonably long period of time, prevent-ing too many forced changeovers. We then solve a relevantoptimization problem, P2(t), shown below, over these con-figurations. P2(t) is an advanced optimization problem and,as a result, the RCM and ADRB policies that are developedby the baseline heuristic are sophisticated. We are confidentthat, on average, any improvement in objective obtainedthrough P1 over a policy resulting from the baseline heuris-tic will represent a lower bound on the improvement thatwould be observed in practice by implementing a policyobtained from P1.

In making such a claim, one cannot ignore the stochasticenvironment in which this problem is solved. Recall, how-ever, that we consider a short time horizon of a few hours,limiting the uncertainty associated with the data. Moreover,implementation of our methodology should use a rollinghorizon approach, re-solving the MIP every 15 minutes orso, to take into account changing conditions while alwayslooking far enough ahead into the future to avoid simplyrelying on a greedy policy.

The following algorithm computes and simulates thebaseline policy.

Algorithm 11. Set t 2= 1, and let f 2= 0 be the simulated cost,

a� 2= a� the actual demand for arrivals and d� 2= d� theactual demand for departures, ∀ � ∈ T. Also let � be thedefault number of consecutive time periods for which aconfiguration must be available to be considered for selec-tion. In this paper, we let � = 6, corresponding to one hour.

2. If Kt 6= �, then go to Step 3; else, no demand isserved at this time period, so update the simulated costand actual demand by setting f 2= f + ctat + qtdt , at+1 2=at+1 + at , dt+1 2= dt+1 + dt , and t 2= t + 1. If t ¶ T , thengo to Step 2; else, Stop.

3. Let Kt =⋂�−1

h=0 Kt+h be the set of configurations eli-gible for selection. If Kt = �, then set � 2= �− 1 and goto Step 3; else, set � 2= 6 and go to Step 4.

4. Given the state of the system, choose the configura-tion to operate in by solving the MIP P2(t) if t = 1, orP2(t + 1) otherwise:

P24t52 minT∑

�=t

4c�u� + q�v�5

s/t0 u� +∑

k∈Kt

yk� = a�1 ∀ � ∈ 8t1 0 0 0 1 T 91 (13)

v� +∑

k∈Kt

xk� = d�1 ∀ � ∈ 8t1 0 0 0 1 T 91 (14)

�jkyk� −�jkxk� −�jkzk� ¶ 01

∀ j ∈Jtk1 ∀k ∈ Kt1 ∀ � ∈ 8t1 0 0 0 1 T 91 (15)

zk� − �k ¶ 01 ∀k ∈ Kt1

∀ � ∈ 8t1 0 0 0 1 T 91 (16)∑

k∈Kt

�k ¶ 11 (17)

zk� ∈ 801191 ∀k ∈ Kt1

∀ � ∈ 8t1 0 0 0 1 T 91 (18)

u�1 v� ∈�+1 ∀ � ∈ 8t1 0 0 0 1 T 91 (19)

yk� ¾ 01 ∀k ∈ Kt1∀ � ∈ 8t1 0 0 0 1 T 91 (20)

xk� ¾ 01 ∀k ∈ Kt1∀ � ∈ 8t1 0 0 0 1 T 90 (21)

Note that Constraints (16) and (17) mean that onlyone configuration can be chosen for the entire time hori-zon t1 0 0 0 1 T . Furthermore, only scheduled demand is

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


considered for any given time period, i.e., the backloggingof demand is not taken into account in Constraints (13) and(14), unlike in the related Constraints (1) and (2) of P1.The underlying rationale is that working with scheduleddemand, rather than going through the difficult process ofcalculating actual demand, more closely resembles the real-world decision-making approach of controllers (as shownin the example problem of §1.1, the effect of controllers’decisions on future demand might be very hard to predict,hence this is a good way to model their decision process).

Hence P2(t) chooses the single best configurationaccording to scheduled demand over the remaining timehorizon. We let k∗ be this configuration, i.e., such that�∗k∗ = 1, where È∗ is the optimal vector � of P2(t), and lett′ = min8�2 � > t1 k∗ yK�9 be the first time at which con-figuration k∗ becomes unavailable (note that this will beT + 1 if k∗ is available for the rest of the time horizon).We will operate in configuration k∗ until time t′ − 1.

5. If t 6= 1, then let � 2= 0, preventing demand frombeing served at time t through Inequalities (24) inP3(k∗1 �1 t1 t′) below; else, let � 2= 1.

6. Now simulate the effect of this choice of k∗ for theinterval 6t1 t′ −17. Solve the following MIP, P3(k∗1 �1 t1 t′):

P34k∗1Ä1 t1 t′52

mint′−1∑

�=t

4c�u� + q�v�5

s/t0 u� − u�−1 + y� = a�1 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (22)

v� − v�−1 + x� = d�1 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (23)

�jk∗yt −�jk∗xt − �0�jk∗zt ¶ 01 ∀ j ∈Jtk∗1 (24)

Figure 7. Test Problem 2, with four configurations, 11 RCCE, and 30 time intervals.

1

2

141312111098

Departures

765432100

1

2

3

4

5

6

7

Arr

ival

s 8

9

10

11

12

13

14

3

4

1, 2

2, 2

3, 2

4, 2

1, 3

2, 3

1, 4

Demand

�jk∗yk∗� −�jk∗x� −�jk∗z� ¶ 01

∀ j ∈J�k∗1 ∀ � ∈ 8t + 11 0 0 0 1 t′ − 191 (25)

z� ∈ 801191 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (26)

u�1 v� ∈�+1 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (27)

y� ¾ 01 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (28)

x� ¾ 01 ∀ � ∈ 8t1 0 0 0 1 t′ − 191 (29)

ut−11 vt−1 = 00 (30)

The above MIP simply fixes the configuration that isoperated over the time interval 6t1 t′ − 17 to be k∗, andsimulates the effect of this decision, forcing zero demandto be served at time t if �= 0.

7. Update the simulated cost and actual demand by set-ting f 2= f +

∑t′−1�=t 4c�u

∗� + q�v

∗�5, at′ 2= at′ + u∗

t′−1, anddt′ 2= dt′ + v∗

t′−1.8. Set t 2= t′. If t ¶ T , then go to Step 2.9. Stop. The baseline value is f .

1.3. Computational Results for Larger Problems

Now we introduce several problems of realistic size, shownin Figures 7, 8, and 9. Note that the set of RCCE in Prob-lem 4 is a superset of that in Problem 3, which is itself asuperset of that in Problem 2. Each problem has 30 timeintervals, representing 24 10-minute intervals, and 6 inter-vals at the end with zero demand. Table 1 outlines theavailability of configurations. In addition, when a configu-ration is available, some of its RCCE might be unavailable,but these details are omitted for brevity.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Figure 8. Test Problem 3, with eight configurations, 22 RCCE, and 30 time intervals.

14 151312111098

Departures76543210

0

1

2

3

4

5

6

7Arr

ival

s

8

9

10

11

12

13

14

15 123456781, 22, 23, 24, 25, 26, 27, 28, 21, 32, 35, 36, 31, 47, 4Demand

Table 2 shows that computation times were in all casesless than three seconds, even for the largest of problems.Indeed, another indication of the strength of the model,also shown in the table, is the small difference between theoptimal objective value of P1, Z∗

IP and the optimal value ofthe LP relaxation of P1, Z∗

LP .We now compare the policies and objective values

obtained from P1 to those of the baseline heuristic. Theresults are presented in Tables 4 and 5. Note that we have

Figure 9. Test Problem 4, with 12 configurations, 33 RCCE, and 30 time intervals.

Departures

15141312

12

13

14

15

11

11

10

10

9

9

8

8

Arr

ival

s

7

7

6

6

5

5

4

4

3

3

2

2

1

1

00

introduced further examples 5–7, with 9, 8, and 9 config-urations, respectively, whose detailed description we omitfor brevity. We do, however, describe the availability ofthese RCCE in Table 3.

It can be observed that in many cases our model far out-performs the baseline approach, by margins of 10%–50%.This can be explained by the cumulative nature of thedemand in the system, which amplifies the negative effectof any suboptimal decisions made early in the time horizon.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 1. Availability of configurations for Prob-lems 2–4.

Problem Configurations that are unavailable

2 1 at t = 20; 2 on 6191217; 3 on 861121259;4 on 85161121131259

3 1 and 5 on 6161177; 2 and 6 on 6151177;3 and 7 on 841279; 4 and 8 on 84151279

4 1, 5, 9, and 11 on 6161177; 2 and 6 on 6151177;3 and 7 on 841279; 4, 8, and 12 on 84151279;10 on 841151279

However, we also note that in one instance where the poli-cies are similar the improvement is only 2%.

A small difference could also be expected when all con-figurations are available at all time periods, because in thiscase the baseline heuristic would produce a policy that usesa single configuration, and P1 would also be likely to do thesame despite its slight difference from P2(1). However, inthis case the heuristic almost becomes a complete optimiza-tion. Furthermore, this case is an uninteresting one becausein practice many configurations often cannot be operatedfor prolonged periods of time. This problem becomes mostinteresting when some configurations become unavailableand we must determine whether to choose the one thatis best now, even though it will become unavailable at alater point, or to choose some other configuration that willremain available for a longer period of time. In any case, itis clear that the potential benefit from the implementationof this method could be significant. Overall, we expect tosee an average improvement of the order of at least 10% inpractice, given the results presented here, and the advancednature of our baseline heuristic, which involves solving arelevant optimization problem.

2. Multiple Changeover TimesWe now extend the model to allow different changeovertimes belonging to a finite set of the form 8iC2 i =

11 0 0 0 1 �9. We must first discretize the time horizon further,such that the length of one time interval is equal to theminimum changeover time C. For a given problem solvedby P1, this will increase the number of time intervals inthe time horizon and scale down the RCCE, both by a fac-tor of �. Note that the new RCCE will then not necessar-ily define the convex hull of the feasible operating points,hence all such RCCE must be rounded so that they do

Table 2. Computation times for P1 and a measure ofthe strength of the formulation.

Computation % difference betweenProblem time (s) Z∗

IP Z∗LP Z∗

IP and Z∗LP

2 0.1 6,148 5,681.4 7063 1.1 4,752 3,376.4 28094 2.6 3,330 2,788.7 1602

Table 3. Availability of configurations for prob-lems 5–7.

Problem Configurations that are unavailable

5 1 at 10, 17, and 18; 2 at 21; 5 at 11, 20, and 21;6 on 6101127, at 20 and 21; 9 at 21

6 3, 4, 5, 6, 7, and 8 on 6101117; 1, 2, 3, 4, 5, and 6on 6201217

7 2, 3, and 4 on 6131147; 6, 7, and 8 at 25

indeed satisfy this requirement. However, we cannot makethe discretization too fine because the rounding error ofthese RCCE would then become significant.

To model changeover times for cases in which one timeinterval now becomes two, we let K′ = 84k11 k251 4k31 k4510 0 0 1 4km−11 km59 be the set of ordered pairs represent-ing changeovers that take two time intervals, where ki ∈⋃

t∈TKt , ∀ i ∈ 811 0 0 0 1m9. We then add the following con-straints, of which there are at most m4T − 25/2, to themodel

zk1 t−2 +∑

k′ 2 4k1 k′5∈K′

zk′t ¶ 11

∀k s.t. ∃4k1 k′5 ∈K′1 ∀ t ∈T\811290 (31)

For three changeover times, which for all practical pur-poses is the most that would be necessary, we first triplethe number of time intervals in the time horizon and thenadd the constraints for those changeovers that take two timeintervals, exactly as in (31) above. In addition, we let K′′ =

84k11 k251 4k31 k451 0 0 0 1 4kn−11 kn59 be the set of ordered pairsrepresenting changeovers that take three time intervals. Wethen add the following constraints to the model:

zk1 t−3 +∑

k′ 2 4k1 k′5∈K′′

zk′t ¶ 11

∀k s.t. ∃4k1 k′5 ∈K′′1 ∀ t ∈T\81121390 (32)

Finally, we can again add valid inequalities in the spiritof Inequalities (11), corresponding to both Constraints (31)and (32).

2.1. Size of the Model

We now present in Table 6 the effect of allowing multiplechangeover times on the size of the model. To simplify,

Table 4. Comparison of solutions from the model P1with baseline solutions.

P1 solution BaselineProblem value solution value % difference

2 61148 61964 13033 41752 71180 51014 31330 31396 2005 81462 121510 47086 81826 101504 19007 101078 111638 1505

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 5. Comparison of the optimal policies and the baseline policies.

Problem P1 policy Baseline policy

2 1 on 611177; 4 on 6191247; 3 on 6261307 1 on 611197; 1 on 62113073 1 on 61157; 4 on 671267 6 on 611147; 4 on 6161267; 4 on 62813074 11 on 611157; 10 on 6171267 11 on 611157; 4 on 61712675 9 on 611197; 1 on 6211307 1 on 61197; 1 on 6111167; 4 on 61813076 2 on 611197; 8 on 6211307 8 on 61197; 2 on 6111197; 7 on 62113077 2 on 611127; 6 on 6141237; 5 on 6251307 3 on 611127; 8 on 6141247; 8 on 6261307

Table 6. Effect of the number of changeover times on the size of the model.

Integer Continuous Upper bound onChangeover times variables variables number of constraints

1 T 42 +K5 (420) 2KT (720) T 43 + J + JK + 4K5 (2,700)2 2T 42 +K5 (840) 4KT (1,440) 2T 43 + J + JK + 4K + 2K25 (22,700)3 3T 42 +K5 (1,260) 6KT (2,160) 3T 43 + J + JK + 4K + 4K25 (60,000)

Notes. T is the length of the time horizon under a single changeover time. Figures in parentheses correspond approx-imately to the largest problem considered in §1.3, with J = 3, K = 12, T = 30. Valid inequalities are included in thesecounts.

we assume that ∀k ∈ Kt , ∀ t ∈ T, we have �Kt� = K and�Jt

k� = �J′t� = J .

Note that in the worst case, allowing for threechangeover times increases the number of constraints by afactor of 20, but the model is still not excessively large.This case occurs when there is a single changeover that isabout a third of the length of all others.

2.2. Computational Results

In the interest of estimating computation times for thecase of multiple changeover times, as well as gainingsome insight into the effect of this modeling adjustmenton optimal solutions, we next modify the three problemsfrom §1.3. We first consider the case of two changeovertimes, allowing the changeovers between two configura-tions to take half the time of all other changeovers. Theseshorter changeovers are indicated in Table 7, along with thecorresponding optimal policies and computation times.

From Table 7, it can be seen that computation timesincreased by an order of magnitude over those of Table 2in §1.3 but were still very short. Furthermore, different opti-mal policies are obtained that favor the shorter changeovers,compared to those of Table 5 in §1.3. In particular, the opti-mal configuration sequence for Problem 2 changes from4114135 to 4113115, while that for Problem 3 changesfrom 41145 to 42135, and that for Problem 4 from 4111105

Table 7. Policies and computation times for twochangeover times.

ShortProblem changeovers Optimal policy Time (s)

2′ 81139 1 on 611357, 3 on 6371487, 0051 on 6501567

3′ 82139 2 on 61177, 3 on 691517 24024′ 841119 11 on 611187, 4 on 6201507 409

to 411145, each clearly taking advantage of the shorterchangeover times, as might be expected.

Next we modify the problems from §1.3 in a similar wayto allow for three different changeover times. The modifi-cations and results are shown in Table 8. Again, the opti-mal policies change due to the different changeover timesbetween certain configurations. For example, comparingwith Table 7, the optimal configuration sequence for Prob-lem 2 changes from 4113115 to 4113125, given the shortchangeover between configurations 2 and 3. Furthermore,the long changeover between configurations 1 and 4 rulesout the optimal solution 4114135 to the original problem,shown in Table 5, §1.3.

While it is clear that computation times are significantlylonger than in the case of one or two changeover times,these results indicate that optimal solutions are obtainedquickly enough to warrant successful implementation inpractice, even for the largest problems.

3. Optimizing Over a Metroplex ofAirports

We now extend the problem to consider multiple airportsoperating within close proximity of one another. This is

Table 8. Policies and computation times for three dif-ferent changeover times.

Problem Short Long Optimal policy Time (s)

2′′ 82139 81149 1 on 611527, 3 on 6551727, 2042 on 6741847

3′′ 83149 81149 2 on 611167, 4 on 6191777 122014′′ 841109 8101119 11 on 611277, 4 on 6301767 4400

Notes. The “Short” column indicates configuration pairs requiring asingle time interval for a changeover, whereas the pairs in the “Long”column require three time intervals.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


often referred to as a metroplex of airports. In this caseit is not sufficient, in general, to optimize each airportseparately, because one might end up with an infeasiblesolution for the system as a whole due to the interactionsbetween arrivals and departures at the different airports.This relationship has been studied recently in Bonnefoy andHansman (2005) and Bonnefoy (2008). In particular, thecapacity of a metroplex as a system will in general be lowerthan the sum of the capacities of its individual airports.

Furthermore, the use of a given configuration at one air-port will often impact the range of configurations that canbe used simultaneously at neighboring airports within themetroplex. Consider, for example, the New York metro-plex consisting of the airports of Newark (EWR), Kennedy(JFK), LaGuardia (LGA), Islip (ISP), and Teterboro (TEB).One of the many instances of the interactions mentionedabove is that when JFK operates with landings on Runway13L in IMC, LGA must also use its Runway 13 for landingsand must also coordinate departures on either its Runway 4or its Runway 13 with JFK (New York ARTCC 2008).

3.1. Multiple Airport Mixed IntegerProgramming Model

In extending the mixed integer programming model to thiscase, an extra index is added to each variable, correspond-ing to the relevant airport p ∈ P, and the constraints anddefinitions of sets and parameters are modified accordingly.In addition, we return to the original assumption of constantchangeover times across all airports. The resulting mixedinteger program is presented below and can be viewed as alarger version of the single airport case, with several cou-pling constraints added.

P42 min∑

p∈P

∑

t∈T

4cptupt + qptvpt5

s/t0 upt − up1 t−1 +∑

k∈Kpt

ypkt = apt1

∀p ∈P1 ∀ t ∈T1 (33)

vpt − vp1 t−1 +∑

k∈Kpt

xpkt = dpt1

∀p ∈P1 ∀ t ∈T1 (34)

�pjkypkt −�pjkxpkt −�pjkzpkt ¶ 01

∀p ∈P1 ∀ j ∈Jtpk1 ∀k ∈Kpt1 ∀ t ∈T1 (35)

� ′

pjt

∑

k∈Kpt

ypkt −�′

pjt

∑

k′∈Kpt

xpk′t ¶ �′

pjt1

∀p ∈P1 ∀ j ∈J′

pt1 ∀ t ∈T1 (36)∑

k∈Kpt

zpkt ¶ 11 ∀p ∈P1 ∀ t ∈T1 (37)

zpkt +∑

k′∈Kpt−1\8k9

zpk′1 t−1 ¶ 11

∀p ∈P1 ∀k ∈Kpt1 ∀ t ∈T\8191 (38)

zpk1 t−1 +∑

k′∈Kpt\8k9

zpk′t ¶ 11

∀p ∈P1 ∀k ∈Kpt−11 ∀ t ∈T\8191 (39)

zpkt + zp′k′t ¶ 11 ∀p ∈P1 ∀k ∈Kpt1

∀ 8p′1 k′9 ∈Npk1 ∀ t ∈T1 (40)

� ij

∑

p∈P

∑

k∈Kpt

ypkt − �ij

∑

p′∈P

∑

k′∈Kp′ t

xp′k′t ¶ �ij1

∀ i ∈I1 ∀ j ∈ Ji1 ∀ t ∈T1 (41)

zpkt ∈ 801191 ∀p ∈P1

∀k ∈Kpt1 ∀ t ∈T1 (42)

upt1 vpt ∈�+1 ∀p ∈P1 ∀ t ∈T1 (43)

ypkt ¾ 01 ∀p ∈P1 ∀k ∈Kpt1 ∀ t ∈T1 (44)

xpkt ¾ 01 ∀p ∈P1 ∀k ∈Kpt1 ∀ t ∈T1 (45)

up01 vp0 = 01 ∀p ∈P0 (46)

Addressing first the similarities of P4 with P1, our objec-tive is again to minimize the total cost of delays incurred,where upt and vpt are the number of arrivals and departures,respectively, that go unserved at airport p at time t. Thesevariables are defined by Constraints (33), (34), (43), and(46). Constraints (35) force the operating point, 4xpkt1 ypkt5,given that we operate in the 4p1 k5th RCCE at time t, tolie within this RCCE, while forcing it to zero if zpkt = 0,in which case we do not operate in RCCE 4p1 k5 at time t.Constraints (37) state that, at any time t, we may operatein at most one configuration at each airport, while Con-straints (38) model the cost of changing from one config-uration to another at each airport by enforcing a delay ofone time interval. Finally, for a fixed airport p, Inequalities(36) are identical to Inequalities (12), which were added toP1; and similarly, Inequalities (39) are the same as Inequal-ities (11), overlooking the notational difference of an extrap in the subscripts.

In addition to these constraints, which are in essencethe same in P4 as in P1, we model the interdependencebetween airports in Constraints (40) and (41) as follows.Letting Npk = 88p′

11 k′191 8p

′21 k

′291 0 0 09 be the set of pairs

such that we cannot operate in configuration k at airport pwhile also operating in configuration k′ at airport p′, for any8p′1 k′9 ∈Npk, Constraints (40) state that we cannot operatein configurations at different airports that are incompatible.In Constraints (41), we assume that for each index i insome set I, we have a system-wide capacity envelope ineffect at every time interval that models the extra depen-dence between airports, and let �i

j , �ij , and � i

j be the param-eters of the jth piece of the ith capacity envelope, wherej ∈ Ji. Indeed, the dependence might be much deeperthan this, ideally requiring capacity envelopes for everycombination of airports and configurations. However, suchenvelopes might be difficult to obtain, requiring sophisti-cated computational modeling of the system, whereas the

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


system-wide envelopes above might be obtained throughobserved system data.

3.2. Computational Results

We show here a large example with 5 airports, which have,respectively, 10, 8, 4, 4, and 4 configurations and 26, 20,11, 10, and 12 RCCE, giving a total of 30 configurationsand 79 RCCE. Problem characteristics are varied, and com-putational results are presented in Tables 9 and 10.

In Table 9, when system-wide capacity envelopes, Con-straints (41), are omitted, one can observe that the solutiontimes are highly variable, depending on the characteristicsof the problem. In varying the characteristics, we start bysolving the full problem (i.e., with all configurations andRCCE available and all combinations allowed) and thenmodify the problem so that the optimal solution becomesinfeasible. Proceeding in this manner, one can see that dis-allowing certain combinations of configurations being usedbetween airports reduces computation time, while remov-ing the availability of certain RCCE and configurations atcertain times can greatly increase computation time.

These trends make sense because disallowing combina-tions of configurations between airports reduces problemsize, while progressively removing optimal solutions bychanging the problem data is likely to result in an opti-mal solution that is more difficult to obtain by branch andbound. Note as well that each of the first three rows ofTable 9 closely corresponds to solving five single airportproblems, because there is no coupling between the fiveairports in these instances. This gives further evidence thatthe models we have presented to solve the single airportproblem are tractable.

Table 10 shows computational results once a system-wide RCCE, Constraint (41), has been added. First, oneshould note that no optimal solutions were obtained withinfive minutes, and hence we present the bounds obtained.This is clearly a significantly worsened performance, com-pared to the single-airport case. However, it is encour-aging that good feasible solutions were always obtained

Table 9. Effect of problem characteristics on compu-tation time for the metroplex case with nosystem-wide RCCE.

Some OptimalConfigurations RCCE 88p1k91 8p′1 k′99 objectiveunavailable† unavailable disallowed‡ value Time (s)

No No No 281422 206Yes No No 351120 2305Yes Yes No 361424 7906No No Yes 281422 103Yes No Yes 351120 2703Yes Yes Yes 361862 3003

†41141135, 41191205, 42121135, 42141205, 43131135, 4412155,4412165.

‡8811991 821399, 8821391 841499, 8831291 851499.

.

Table 10. Effect of problem characteristics on thebounds on the optimal objective valueobtained within five minutes for the metro-plex case with a system-wide RCCE.

Some 88p1k91 Bounds onConfigurations RCCE 8p′1 k′99 optimal Optimalityunavailable† unavailable disallowed‡ objective value gap (%)

No No No 483139718319045 006Yes No No 483139718517445 207Yes Yes No 483139718413425 101No No Yes 483139718318965 006Yes No Yes 483139718512065 201Yes Yes Yes 483139718412205 100

†41141135, 41191205, 42121135, 42141205, 43131135, 4412155, 4412165.‡8811991 821399, 8821391 841499, 8831291 851499.

within one minute and that the best solutions obtained afterfive minutes typically were within about 2% of optimality.In addition, observe that in this example the system-wideRCCE is quite restrictive given the increase in the opti-mal objective values that can be inferred from comparingTables 9 and 10.

It is evident that the addition of a system-wide RCCEhas a significant effect on computation time, and that as aresult the metroplex formulation is less effective than thesingle airport formulation. Nevertheless, good solutions areobtained within a practical amount of time.

4. Further ExtensionsWe consider next extensions to the basic model that addressissues occasionally arising in practice. Suppose that thereare certain environmental constraints on certain configura-tions, such as federal or local regulations (or, in many cases,local “letters of understanding”) governing airport opera-tions. For example, the use of a configuration that leadsto departures taking off over a residential area might beprohibited or discouraged during certain times of the day.We show here that the model can be easily extended toaccommodate such restrictions.

Consider, for example, the following cases for the singleairport problem:

1. The maximum total operating time in configuration kis S1 time intervals in any continuous period of lengthS2 >S1. We add the following T + 1 − S2 constraints:

∑

t∈Q

zkt ¶ S11 ∀ t′ ∈ 81121 0 0 0 1 T + 1 − S291

where Q= 8t′1 t′ + 11 0 0 0 1 t′ + S2 − 19⊂T0 (47)

2. Once we operate in configuration k, and then stop, wemay not resume operation in this configuration until it hasbeen inoperative for S time intervals. Assuming that thereare at least S time intervals remaining after some time t,we want zkt = 11 zk1 t+1 = 0 ⇒ zk1 t+21 zk1 t+31 0 0 0 1 zk1 t+S = 0.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


We add 2T − 4 variables, wt and �t , and the following3T − 6 + 4T − S54S − 15+ 4S − 254S − 15/2 constraints toour model:

wt ¾ 01 ∀ t ∈ 81121 0 0 0 1 T − 291 (48)

wt ¾ zkt − zk1 t+11 ∀ t ∈ 81121 0 0 0 1 T − 291 (49)

�t ¾ zk1 t+j+11 ∀ t ∈ 81121 0 0 0 1 T − 291 ∀ j ∈ 81121 0 0 0 1 s′91

where s′= min8S − 11 T − t − 191 (50)

�t +wt ¶ 11 ∀ t ∈ 81121 0 0 0 1 T − 290 (51)

Note from Constraints (49) that wt ¾ 1 if we operate inthe kth configuration at time t but not at time t + 1, andfrom Constraints (50) that �t ¾ 1 if we operate in configu-ration k in any of the s′ time intervals after interval t + 1.Constraints (51) then state that at most one of these eventscan occur.

Given that such extensions would normally complicatethe decision-making process significantly, the ease withwhich the MIP models can incorporate them is an impor-tant benefit.

5. ConclusionsWe have presented a strong mixed integer programmingformulation to solve the single airport RCM and ADRBproblems. Evidence provided in the form of computa-tional results on realistic problem data indicates that ourmodel is capable of solving real-world instances to opti-mality quickly enough to be implemented. To obtain alower bound on the expected benefits from our optimiza-tion approach, a sophisticated optimization-based heuristicwas developed that reveals the potential cost savings to beindeed significant. It was also shown that a number of addi-tional potential constraints and local considerations can beincorporated into the models with little difficulty. Finally,we have also proposed an extension of the model to opti-mize over a metroplex of airports. Although the resultsreported herein are encouraging, this is a topic that lendsitself to much future work.

We have now started, in association with the MITLincoln Laboratory, a process that might eventually leadto field implementation of our approach. As a first step,the optimization models would be operated “offline” for

one or more selected airports and their recommendations,and attendant delays would be compared with the actualsequences selected by the FAA and the observed delaysto arrivals and departures. If successful, we envisage thatin the long run the approach will be implemented usinga rolling horizon, with the solution being updated at rela-tively short intervals (e.g., every 30 minutes) to compensatefor the uncertainty in some of the data.

AcknowledgmentsThe authors thank Bill Moser of MIT Lincoln Laboratoryand Tom Reynolds of MIT Partnership for Air Transporta-tion Noise and Emissions Reduction and of MIT LincolnLaboratory for helpful discussions. This work was partiallysupported by NSF grant EFRI-0735905 and the MIT Lin-coln Laboratory.

ReferencesBonnefoy, P. 2008. Scalability of the air transportation system and devel-

opment of multi-airport systems: A worldwide perspective. Ph.D. the-sis, Engineering Systems Division, Massachusetts Institute of Tech-nology, Cambridge, MA.

Bonnefoy, P., R. J. Hansman. 2005. Emergence and impact of secondaryairports in the United States. FAA/Eurocontrol Air Traffic Manage-ment Conf. (ATM), Baltimore.

de Neufville, R., A. Odoni. 2003. Airport Systems: Planning, Design, andManagement. McGraw-Hill, New York.

Gilbo, E. P. 1993. Airport capacity: Representation, estimation, optimiza-tion. IEEE Trans. Control Systems Tech. 1(3) 144–154.

Gilbo, E. P. 1997. Optimizing airport capacity utilization in air trafficflow management subject to constraints at arrival and departure fixes.IEEE Trans. Control Systems Tech. 5(5) 490–503.

Gilbo, E. P., K. W. Howard. 2000. Collaborative optimization of airportarrival and departure traffic flow management strategies for CDM.3rd USA/Europe Air Traffic Management R&D Seminar, Naples,Italy.

Hall, W. D. 1999. Efficient capacity allocation in a collaborative airtransportation system. Ph.D. thesis, Operations Research Center,Massachusetts Institute of Technology, Cambridge, MA.

Kellner, S., D. Kösters. 2008. Contribution to the ATMAP project. Work-ing paper, Institute of Transport Science, RWTH Aachen University,Aachen, Germany.

New York ARTCC. 2008. N90 TRACON—LaGuardia standard operatingprocedures. New York ARTCC, New York.

Stamatopoulos, M. A., K. G. Zografos, A. R. Odoni. 2004. A decision sup-port system for airport strategic planning. Transportation Res. Part C12 91–117.

Swedish, W. J. 1981. Upgraded FAA airfield capacity model. Report FAA-EM-81-1, Vols. 1 and 2, The MITRE Corporation, McLean, VA.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

Documents

Optimal Selection of Airport Runway Conﬁgurationsdbertsim/papers...Bertsimas, Frankovich, and Odoni: Optimal Selection of Airport Runway Conﬁgurations 1408 Operations Research